Full-Stack & Applied AI Engineer

Seif Osman — Full-Stack & Applied AI Engineer

open to roles— NYCStaten Island, NY

Store tech by day. Production AI by night.

5+ years on the team that keeps POS and store networks running across an 890-store fleet. In October 2026 I specified, directed and launched an AI phone receptionist and booking platform for a real client.

View work Resume Contact

890 stores supported200+ stores set up A→Z1 AI system in production10+ techs trained

Five years on the first-call line for 890 stores. Then I started shipping software of my own.

I'm a Computer Science graduate (CSI / CUNY, 2026) with 5+ years on the store-technology team at Rainbow Apparel, where we build, deploy and support POS systems, Cisco store networks, connectivity and peripherals across an 890-store environment. I'm one of the people a store calls when a register is down, and I train incoming technicians.

Each light stands for one of the 890 stores my team supports. The four rings are the work I do across them: choose one to read about it.

An illustration of scale, not a map or live data.

Skip ahead to my work

Outside the day job I build and ship software end to end: a production AI phone receptionist and booking platform (live since October 2026), a 100k-line invoicing and inventory platform, a measured media-automation pipeline, and local-first and security tools. All of it is tested, most of it is CI-gated, and the 2026 work was AI-assisted under my direction — which I say plainly.

Choose a project to see what it is.

From a missed call to a booked job

A small service business went from missed calls to a system that answers every call, collects the caller’s details, books within the business’s rules and confirms by text. The client is unnamed pending consent.

Choose a step to follow the call.

Illustration · scripted example, not a live system

The same discipline, both ways

Five years on the first-call line for a nationwide fleet: a dead register is a store losing money. I bring the same discipline to my own systems — restart-safe state, database-enforced rules, and every incident recorded and closed with a fix commit.

Applied AITypeScriptPostgreSQLPythonPOS / Store TechCisco Networks

Name: Seif Osman
Role: Full-Stack & Applied AI Engineer
Degree: B.S. Computer Science · CUNY
Experience: 5+ years · 890 stores
Builds: 14 featured · 1 in production
Status: open to roles

01 Selected work

Things I've shipped.

14 featured builds — a production AI system, business platforms, AI tooling and security tools. Five cards; scroll to deal them.

Illustration · animated

Production AI system
AI Receptionist & Booking Platform
AI phone answering and booking, with SMS confirmations.
live since Oct 2026
Case study
Real screens · synthetic data

Business platform
Asset Vault
Invoicing, inventory transfers and PDF generation in one platform.
368 unit · 111 real-DB tests
Case study
Illustration · animated

Media automation
PeakClips Pipeline
Turns long videos into vertical clips with highlights and captions.
analysis gate 590 s → 17 s
Case study
Real captures · sample text

Local-first desktop app
Store Ticket Assistant
Transcribes support calls on the device and drafts the ticket.
292 tests in CI
Case study
Screenshots · synthetic data

The rest of the deck
More projects
Ten more case studies, from packet sniffer to POS.
Browse all 14

Scroll to deal the cardsBrowse all 14

All fourteen projects

Every build with its own case study, in the order they appear on this site.

890 stores, one network

The footprint my team keeps online — registers, Cisco networks, connectivity and peripherals across a nationwide retail fleet. Hover a node to inspect a storeEvery node is a real store (the footprint is real; the telemetry below is a simulation).

Fleet view · simulated

scanning…

890Stores supported

5+Years on the fleet

200+Stores set up A→Z

SIMTelemetry below is simulated

OnlineNeeds attentionResolving ticket

Ops feed · simulated--:--:--

Ring me up

I deploy point-of-sale for a living — so here's one, selling the only thing in stock: me. Every item on the register is a real thing I do — store tech by day, full-stack & security builds by night. Add a few, hit checkout, and it prints your quote.

Seif's Dev Store OPEN Ready · scan an item

★ TODAY'S SPECIAL

Services7 items

↕ scroll for more · press 1–9 to scan

Subtotal$0.00

Tax · NYC 8.875%$0.00

Total$0.00

🔥 Demo sales: 0 orders · $0.00

02 Back of house

Why teams keep me

I ship things that stay shipped.

Production

An AI phone receptionist and booking platform for a real client has been live since October 2026 — gated releases, a recorded rollback floor, and a 15-finding security review in which every high and medium finding was fixed or mitigated.

1,845tests at the last release gate

Read the case study

Downtime is the enemy.

Reliability

10incidents recorded, each closed with a fix commit

Attacker's eye.

Security

$ security+ in progress$ thm / htb active$ wireshark · tcpdump · scapy

I multiply teams.

Team

Trained 10+ IT technicians from zero — I document as I build.

10+techs trained

03 How I build

AI-assisted, evidence-gated

Since 2026 most of my line-by-line code has been generated by an AI pair under my direction. I say so plainly, and I hold the result to a higher standard of proof because of it. Here is what I own, how work gets verified, and how I label a claim.

I ownthe specification, the safety rules, the code review and the go-live call.

Work is verified byreproducing bugs on the real database or device, positive controls beside every denial test, and gates that must fail on the old code.

The AI pairdrafts to my spec. It never approves a deploy, spends money or talks to a customer.

Read the full explanation

Scope, safety rules, review, and the go-live call.

What I own

I write the specification and the acceptance criteria before the build, set the rules the system may never break (no unapproved spending, no invented business facts, no call ever transferred), read the code and the call transcripts, run every real-device test myself, and approve every deploy. The AI pair drafts; I decide what ships and I carry the consequences.

1,845tests passed at the receptionist's release gate, 2026-10-19

See the case study

Reproduce before fixing. Prove the test can fail.

How work is verified

A bug is reproduced on the real database or the real device before anyone fixes it, and the reproduction becomes the regression test. Every denial test gets a positive control beside it. Mocks are checked against the installed library. A new check has to fail on the old code before it counts. Releases pass a zero-failure gate with a recorded rollback floor, and CI gates on type-checks and tests, not on reports.

45real-database access-control tests; 16 of the first 42 failed on the original code

The positive-control story

Every number has a source.

Evidence standards

$ verified reproduced from a repo, CI run or measurement$ doc stated in a project document, not re-run$ self-reported my own statement, consistent across my materials

It drafts. It does not decide.

Where the AI sits

The pair writes code to my specification and I review it. It never approves a deploy, never spends money, never talks to a customer, and its claims are checked the same way mine are: by running the thing.

10incidents recorded, each closed with a fix commit

Two examples. A budget monitor summed a nested billing API and sent a false $60 alert; the fix was a regression test from the real provider snapshot and a standing rule that no computed money figure may alert anyone until it matches the provider's own total. A video analysis gate went from 590 s to 17 s on a 480p proxy, and the flag stayed off until a diff harness showed that 30 of 31 verdicts were identical, the other one stricter and re-checked at full resolution, with zero missed drops.

Read the notes

Career, version-controlled

Five years of commits, four branches — and one open pull request.

feat: join Rainbow Apparel

main · 2021 — now

IT Store Technology — POS Systems & Infrastructure. Build, image and deploy POS/register systems; diagnose failures across applications, Cisco networks, VLANs and endpoints; train incoming technicians.

890 stores · 200+ set up A→Z · 10+ techs trained

feat: inventory & systems analyst

pickle-works · merged 2024

Secured warehouse inventory in Oracle NetSuite — RBAC, an anti-fraud randomized barcode flow, bug-hunting with Oracle engineers.

RBAC access · 0 tolerance for bad data

wip: attacker's eye

security-track · always open

Packet Analyzer (Scapy/Textual), Wireshark & tcpdump, Google Cybersecurity coursework, TryHackMe & HackTheBox labs.

Security+ in progress · THM/HTB active

merge: B.S. Computer Science

edu/bs-cs · merged jun 2026

CSI / CUNY — earned while running the fleet full-time. IEEE member; Robotics Club.

B.S. CS · IEEE · Robotics

feat: ship a production AI system

main · oct 2026

Specified, directed and launched an AI phone receptionist and booking platform for a real client — 470 commits in 30 days, a 15-finding security review (every high and medium finding fixed or mitigated), gated releases with a rollback floor.

470 commits · 52 rounds · 15 findings reviewed, high and medium fixed or mitigated

pull request · open

Merge seif → your-team

Shipped a production AI system (live Oct 2026)
5+ years keeping production retail alive
Ships, documents & trains the team
Security habit — probes his own builds
B.S. Computer Science, June 2026

✓ No conflicts with your codebase

Merge pull request Review the work

Flagship — AI Receptionist

receptionist · live since oct 2026

24/7answers every call

330ms classification, Sep 2026 measurement

1,845tests at the gate

Incoming call · intent: book_appointment

local modelclassified

Address and ZIP read back · confirmed

intakeconfirmed

Slot held · one job at a time

PostgreSQLexclusion constraint

Caller says “never mind” · booking withdrawn

round 52handled

SMS confirmation · carrier delivered

outboxsent

live since Oct 2026

A phone receptionist that books real appointments — and never promises what it can't deliver

The problem. A sole-operator service business was losing jobs to unanswered calls and had no online booking. It needed a receptionist that answers at any hour, books only inside the business's rules, and stays inside a fixed budget.

What I did. Specified, directed and launched the system: TypeScript on Fastify and PostgreSQL 17, Twilio voice and SMS with cloud speech, and a grammar-constrained local Qwen3 model on llama.cpp for intent classification. Scheduling rules live in the database as exclusion constraints; every release passes a zero-failure gate with a recorded rollback floor; a 15-finding security review was run, and every high and medium finding was fixed or mitigated.

The result. Live since October 2026. Intent classification cut from ~4,000 ms to ~330 ms by keeping the model resident (my September measurement; a later benchmark of the production configuration gave a 470 ms median); ten documented incidents resolved before and after launch. This is early production use, so what is proven is engineering readiness, not business outcomes.

1,845tests at the last gate

15security findings reviewed, high and medium fixed or mitigated

46verified migrations

TypeScriptFastify 5PostgreSQL 17Twiliollama.cpp · Qwen3

Read the case study

5+years · Store IT & POS

890stores supported

1AI system in production

11repos · CI green

Credentials B.S. Computer Science · CUNY 2026 Google PM Professional CodePath